a dynamic discrete-continuous choice model of car
TRANSCRIPT
Aurélie Glerum EPFL
Emma Frejinger Université de Montréal
Anders Karlström KTH
Muriel Beser Hugosson KTH
Michel Bierlaire EPFL
A DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL OF CAR OWNERSHIP AND USAGE
13th Swiss Transport Research Conference
25th April 2013
OUTLINE 2
• Introduction
• Background and data
• The dynamic discrete-continuous choice modeling framework • Assumptions • Definition of the components • Solving the dynamic programming problem • Model estimation
• Illustration of model application
• Conclusion and future works
INTRODUCTION 3
Aim of the research project: • Model dynamics of car transactions, usage and choice of fuel type in the
Swedish car fleet
• Motivations • Governmental policies:
• Goals of reducing carbon emissions • Technology changes:
• Increase of alternative-fuel vehicles • Changes in the supply
• Company cars: represent important share of new car sales
INTRODUCTION 4
Current literature on car ownership and usage modeling:
• Car ownership models in transportation literature: • Mostly static models:
• Main drawback: do not account for forward-looking behavior • Important aspect to account for since car is a durable good
• Econometric literature: dynamic programming (DP) models + discrete choice models (DCM)
• Recently, dynamic discrete choice models (DDCM) starting to be applied in transportation field (Cirillo and Xu, 2011; Schiraldi, 2011)
• Joint models of car ownership and usage:
• Early references: e.g. duration models and regression techniques for car holding duration and usage (De Jong, 1996)
• Dynamic programming mixed logit (DPMXL) (Schjerning, 2007) used to model car ownership, type of car and usage (Munk-Nielsen, 2012)
• Discrete-continuous model of vehicle choice and usage based on register data (Gillingham, 2012)
Research issues:
• Car are durable goods Need to account for forward-looking behavior of individuals
• Difficulty of modeling a discrete-continuous choice
• Many models focus on individual decisions, but choices regarding
car ownership and usage made at household level
INTRODUCTION 5
Research issues:
• Car are durable goods Need to account for forward-looking behavior of individuals
• Difficulty of modeling a discrete-continuous choice
• Many models focus on individual decisions, but choices regarding
car ownership and usage made at household level
Proposed methodology:
• Attempt to address these issues by applying dynamic discrete-continuous choice model (DDCCM)
• Large register data of all individuals and cars in Sweden
INTRODUCTION 6
BACKGROUND AND DATA 7
Register data of Swedish population and car fleet: • Data from 1998 to 2008
• All individuals
• Individual information: socio-economic information on car holder (age, gender, income, home/work location, employment status/sector, etc.)
• Household information: composition (families with children and married couples)
• All vehicles
• Privately-owned cars, cars from privately-owned company and company cars
• Vehicle characteristics (make, model, fuel consumption, fuel type, age) • Annual mileage from odometer readings • Car bought new or second-hand
Aim of the project: • Model simultaneously car ownership, usage and fuel type.
In details: model simultaneous choice of
8
Transaction type
Private/ company car – car c
Fuel type – car c
New/2nd hand – car c
Annual milage – car c
×
×
×
×
# cars
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL
Aim of the project: • Model simultaneously car ownership, usage and fuel type.
In details: model simultaneous choice of
9
Transaction type
Private/ company car – car c
Fuel type – car c
New/2nd hand – car c
Annual milage – car c
×
×
×
×
# cars
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL
Discrete variables
Aim of the project: • Model simultaneously car ownership, usage and fuel type.
In details: model simultaneous choice of
10
Transaction type
Private/ company car – car c
Fuel type – car c
New/2nd hand – car c
Annual milage – car c
×
×
×
×
# cars
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL
Continuous variables
11
Motivations for discrete-continous vs discrete model
• Mileage variable(s) are continuous: lose information by
discretizing it.
• In a discrete-continuous approach: If choice of mileage conditional on the discrete choice Reduction of size of the discrete action space
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL
12
• Decisions at household level: up to 2 cars in household
• Strategic choice of:
• Transaction • Type(s) of ownership (company vs private car) • Fuel type(s) • Car state(s) (new vs 2nd-hand)
Account for forward-looking behavior of households
• Myopic choice of:
• Annual mileage(s)
• Choice of mileage conditional on choice of discrete variables
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL ASSUMPTIONS
• Agent: household
• Time step t: year
• State space S
Age – 1st car
Age – 2nd car
Private/ company car – 1st car
Private/ company car – 2nd car
Fuel type – 1st car
Fuel type – 2nd car
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL 13
DEFINITION OF THE COMPONENTS
14
• Agent: household
• Time step t: year
• State space S
Age – 1st car
Age – 2nd car
Private/ company car – 1st car
Private/ company car – 2nd car
Fuel type – 1st car
Fuel type – 2nd car
2-car households
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
15
• Agent: household
• Time step t: year
• State space S
Age – 1st car
Age – 2nd car
Private/ company car – 1st car
Private/ company car – 2nd car
Fuel type – 1st car
Fuel type – 2nd car
1-car households
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
16
• Agent: household
• Time step t: year
• State space S
Age – 1st car
Age – 2nd car
Private/ company car – 1st car
Private/ company car – 2nd car
Fuel type – 1st car
Fuel type – 2nd car
0-car households
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
17
• Agent: household
• Time step t: year
• State space S
Age – 1st car
Age – 2nd car
Private/ company car – 1st car
Private/ company car – 2nd car
Fuel type – 1st car
Fuel type – 2nd car
= 1407 relatively small state space
= 6
= 4
= 4
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
18
• Action space A
Transaction type
Mileage - 2nd car
Private/ company car – 2nd car
Fuel type – 1st car
Fuel type – 2nd car
New/ 2nd hand – 1st car
Mileage – 1st car
Private/ company car – 1st car
New/ 2nd hand – 2nd car
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
19
• Action space A
Transaction type: details
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
20
• Action space A
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
21
• Action space A
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL
Size of the discrete part of the action space assuming: • 3 levels company car • 3 levels fuel type • 2 levels new/2nd hand
Maximum size of action space << for DDCM
DEFINITION OF THE COMPONENTS
22
• Transition rule: deterministic rule: each state 𝑠𝑡+1 can be inferred exactly once 𝑠𝑡 and 𝑎𝑡 are known.
Example: If 𝑠𝑡 = [2,1,2,0,0,0] and 𝑎𝑡 = 1,12′000,0,0,0,0,3,0,0 then 𝑠𝑡+1 = 3,1,2,0,3,0 .
2 years
private car
diesel
increase 1
mileage 1st car
company car
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
23
• Instantaneous utility function:
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
24
• Instantaneous utility function:
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
Deterministic term
Random term
25
• Instantaneous utility function:
Assume additive deterministic utility for simplicity (see also Munk-Nielsen, 2012):
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
Deterministic term
Random term
26
• Instantaneous utility function:
Assume additive deterministic utility for simplicity (see also Munk-Nielsen, 2012):
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
Deterministic term
Random term
Utility for discrete actions
Utility for continuous actions
27
• Instantaneous utility function:
Assume additive deterministic utility for simplicity (see also Munk-Nielsen, 2012):
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
Deterministic term
Random term
Utility for discrete actions
Utility for continuous actions
28
• Instantaneous utility function
• Utility for continuous actions: Constant elasticity of substitution (CES) utility function:
• Captures substitution patterns between the choice of both annual driving distances
• ρ is elasticity of substitution
• Randomness introduced in
• 𝑥𝑡 contains price of fuel, car consumption and other socio-economic characteristics
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL DEFINITION OF THE COMPONENTS
29
1. Finding the optimal value(s) of annual mileage conditional on the discrete choices
2. Solving the Bellman equation
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL SOLVING THE DYNAMIC PROGRAMMING PROBLEM
30
Finding the optimal value(s) of mileage
• Maximization of the continuous utility:
s.t. • Find analytical solutions: and
• Optimal continuous utility
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL SOLVING THE DYNAMIC PROGRAMMING PROBLEM
31
Solving the Bellman equation
• Logsum formula used in the completely discrete case (DDCM)
(Aguirregabiria and Mira, 2010; Cirillo and Xu, 2011)
• Logsum can be applied here given the key assumptions: • Choice of mileage(s) is conditional on discrete actions • Choice of mileage(s) is myopic
• Iterate on Bellman equation to find integrated value function 𝑽�
DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL SOLVING THE DYNAMIC PROGRAMMING PROBLEM
32
• Parameters obtained by maximizing likelihood:
• Optimization algorithm is Rust’s nested fixed point algorithm (NFXP) (Rust, 1987): • Outer optimization algorithm: search algorithm to obtain parameters
maximizing likelihood
• Inner value iteration algorithm: solves the DP problem for each parameter trial
• Plan to investigate variants of NFXP to speed up computational time
MODEL ESTIMATION DYNAMIC DISCRETE-CONTINUOUS CHOICE MODEL
33
Assumptions for the example:
• Size state space = 651 • Max age = 3 • Company car levels = 3 • Number of fuel types = 3
• Size action space = max 745
• Number of transaction types = 9 • Number of state levels (new/old) = 2
• Utility function contains:
• Transaction cost 𝜏 • Transaction-dependent parameters for age of oldest car
• Parameters of DP problem: • Discount factor β = 0.7 • Stopping criterion 𝜀 = 0.01
postulated
ILLUSTRATION OF MODEL APPLICATION
34
Program:
• Code in C++
• 2 minutes on 20-core server Graph:
• 𝑽� vs age of oldest of cars in
household fleet
ILLUSTRATION OF MODEL APPLICATION
35 ILLUSTRATION OF MODEL APPLICATION
36 ILLUSTRATION OF MODEL APPLICATION
Instantaneous utility
37 ILLUSTRATION OF MODEL APPLICATION
Expected discounted utility
CONCLUSION AND FUTURE WORKS 38
Conclusion:
• Methodology to model choice of car ownership and usage dynamically
• Example of application shows feasibility of approach
Next steps: • Exploratory analysis to specify instantaneous utility
• Model estimation on small sample of synthetic data
• Model estimation on register data
• Scenario testing: • Validation of policy measures taken during the years available in the data
• Test policy measures that are planned to be applied in future years
39
Thanks!